AI & Technology

Active Learning for Systematic Reviews

How machine learning prioritizes likely-relevant records, where workload savings can arise, and why recall must be validated for each review.

By Lumina Editorial Team Published Updated Methodology guide
Lumina live screening queue with AI-ranked papers and reviewer controls
Active learning changes the order of the queue while reviewers retain each inclusion and exclusion decision.

What Is Active Learning?

Active learning is a machine learning approach where the model learns from your screening decisions during screening. As you include or exclude papers, Lumina refreshes the semantic ranking; after enough new decisions, it retrains the classifier and blends that signal into the queue.

1

You Screen

Make a few include/exclude decisions

2

AI Learns

Ranking updates; classifier retrains periodically

3

Re-rank

Remaining papers are re-ordered

4

Evaluate

Validate before changing the stopping point

How It Works Under the Hood

Text Representation

Each paper's title and abstract are converted into a numerical representation (embedding) that captures its semantic meaning.

Tools like Lumina use advanced models (e.g., OpenAI embeddings) that understand research terminology, synonyms, and context.

Relevance Ranking

Your inclusion criteria are converted to the same embedding space. Papers are then ranked by similarity — the more similar to your criteria, the higher they rank.

After each decision, the model also learns from your include/exclude patterns to refine the ranking. In Lumina, relevance feedback can update after a decision while the classifier retrains in batches rather than after every individual label.

Uncertainty Sampling

Some active learning systems also prioritize papers the model is most uncertain about. These are the most informative for learning.

Stopping Criteria

Once the observed yield becomes low, stopping strategies can prompt a review of whether to continue. They do not guarantee that no relevant records remain.

What Does the Evidence Say?

What published evaluations support

📊

van de Schoot et al. (2021) describe ASReview and show that active-learning performance can be evaluated retrospectively on known review datasets. Read the paper.

📊

O'Mara-Eves et al. (2015) found promising uses for text mining, while also emphasizing heterogeneity in methods and evaluation. Read the systematic review.

📊

Practical implication: results from one benchmark do not establish recall for a new review. Dataset size, prevalence, seed records, labeling quality, ranking model, and stopping rule all affect performance.

Active Learning vs. Manual Screening

Manual Screening Active Learning
Order Random or chronological Most relevant first
Workload Screen 100% of papers Potentially fewer, after review-specific validation
Recall ~100% (human error aside) Dataset-dependent; estimate and report uncertainty
Adapts No Yes — learns from each decision

How to Report AI-Assisted Screening

If you use active learning, transparently report:

  • The tool used (e.g., Lumina, ASReview, Rayyan)
  • The stopping criteria applied
  • Total papers vs. papers actually screened
  • Whether dual screening was used alongside AI
  • Any sensitivity analysis or validation performed

Experience Active Learning in Action

Try Lumina's AI screening with a demo dataset and see how active learning re-ranks papers after each decision.